I remember a colleague telling me that he was a bonsai artist, an artist who works with bonsai, and that he received tremendous praise for the work he did with those beautiful pieces, which, strictly speaking, functioned almost like art installations.
He said he was very proud of the results, but also confessed that he felt like a fraud whenever someone asked, “Did you make it?”
The question bothered him because, although there was evident artistic work in every choice, every pruning, every wire used to guide a branch, every restraint placed on growth, and every decision about form, proportion, and negative space, there was a tree there. A real tree. A culture of living cells, with its own metabolism, its own tendencies, its own responses to environmental conditions, and a morphogenesis that he could never control in every detail.
He did not construct every leaf, every root, or every conducting vessel. He did not design the final organism cell by cell. He created conditions, imposed constraints, provided resources… he removed excess, guided tendencies, waited, observed, and intervened again, but he never felt comfortable saying, “I made it.”
More than a bonsai artist, then, he felt like a cultivator.
…and I am talking about this in order to talk about Artificial Intelligence.
There is something profoundly similar in the way we build artificial neural networks. We design architectures, establish training regimes, choose data, define loss functions, regulate learning rates, select examples, and subject a network to enormous quantities of stimulus, error, correction, and renewed exposure. We bear immense responsibility for what emerges from this process, of course, but it would be a very poor description to say that we directly wrote every functional structure that comes to exist within a trained model… not least because we did not.
No one manually programs the concept of irony into a Language Model. No one chooses the exact set of parameters in which the relationship among monarchy, king, power, succession, and legitimacy will reside. No one designs, rule by rule, every circuit that will be used when the model recognizes an analogy, translates an expression, resolves an ambiguity, completes a logical structure, or relates two concepts that had never appeared together in the training data.
We build the architecture and the cultivation regime. What emerges internally results from the interaction among that architecture, the data, the optimization process, and the enormous number of statistical pressures to which we subject the system.
We are not, therefore, the designers of the inferential capacity of that resulting geometry. We built the architecture, selected the data, established training pressures, and created the conditions in which it could emerge; but we did not design inference, we did not program it, and, more importantly, we did not even know that it would emerge. We discovered it.
We did not teach the geometry how to infer. We discovered that certain geometries, when cultivated under particular pressures, learn to do so.
Perhaps that is precisely why certain comparisons between artificial and biological neural networks strike me as increasingly strange. We have rightly discovered that biological neurons are far more complex than the early abstractions used in the history of artificial neural networks. We have discovered that dendrites participate actively in processing, that signal integration can occur in nonlinear ways, that temporality, synaptic location, and cellular dynamics matter, and that the biological unit we long summarized for teaching purposes as a simple integrator is, in fact, an extraordinarily rich physical structure.
All of this is fascinating, but it is fascinating because it teaches us more about biological neurons... not because it establishes, by some automatic consequence, what every possible intelligence must be.
An artificial heart does not need to be subject to all the contingencies and fragilities of a human heart in order to pump blood. It does not need to have undergone embryogenesis, contain living cardiac tissue, manage cellular metabolism, or produce and repair proteins. It does not need to suffer from exactly the same diseases or age in the same way. It can perform a significant part of circulatory function through materials, geometries, energy sources, and maintenance mechanisms that are completely different.
Even so, no one needs to believe that there is a tiny human heart hidden inside the machine in order to understand why we call it an artificial heart.
The word “artificial” already acknowledges a difference in realization. What is preserved is a functional class.
This distinction seems obvious in almost every field of engineering. An airplane does not need to reproduce the anatomy of a bird in order to fly, a submarine does not need to swim like a fish, a camera does not need to reproduce the retina in order to capture and organize light, and a prosthesis does not need to grow from bone tissue in order to bear weight. Engineering is full of solutions that achieve functionally convergent results precisely through morphological paths different from those found by evolution.
Curiously, when we turn to cognition, the observational freedom we acknowledge in other fields mysteriously and frequently disappears from discourse about cognition.
It begins to seem necessary for a machine to exhibit something sufficiently similar to human neurons, human dendrites, human metabolism, a human body, a human learning history, and even human fragilities before the possibility of a given cognitive function is taken seriously.
There is a confusion here between two things that should remain separate: the conditions necessary for a specific biological realization to function and the conditions necessary for a particular function to exist in any possible realization.
A biological neuron is a living cell. This means that an important part of its complexity does not exist because “thinking requires all that complexity,” but because a cell needs to remain a cell. It maintains membranes, regulates electrochemical gradients, synthesizes components, transports substances, manages waste, responds to chemical changes, repairs damage, consumes energy, and participates in a metabolic network without which it would simply cease to exist.
This does not make the cellular mechanisms involved irrelevant to cognition. Many of them participate directly in neural processing, plasticity, signal timing, and synaptic integration. The point is different, and simpler: a biological architecture must perform its function while simultaneously preserving the life of every unit that participates in that function.
Engineering can distribute these responsibilities differently.
An artificial unit does not need to metabolize individually because its physical support is provided at another scale. Energy, heat dissipation, structural integrity, redundancy, storage, and maintenance belong to the broader system in which that unit is realized. Part of what, in biology, must exist within each cell can be moved outside it.
Local simplicity, therefore, does not necessarily imply functional poverty. It may merely signify a different division of responsibilities.
This difference matters because complexity is often used as an argument from authority. One points out that a neuron is biologically far more complex than an artificial unit and, from there, suggests that the latter must be cognitively rudimentary by definition. But the number of processes required to keep a cell alive cannot be indiscriminately added to what we call cognition and then presented as proof of intrinsic cognitive superiority.
What matters is which processes make a difference to the function we are trying to understand.
More than that… which of these differences must be preserved by any other realization capable of achieving a comparable function.
Evolution has never seemed particularly committed to equivalence. Similar functional problems arose in different lineages and were solved by different architectures. Flight arose more than once. Eyes arose more than once. Forms of navigation, chemical perception, echolocation, aquatic locomotion, and countless other capacities were realized by structures with distinct evolutionary histories.
This does not mean that all these solutions are the same. They are not.
It means that a function does not necessarily belong to the morphology in which we first came to know it.
Engineering radicalizes this principle because it does not need to reproduce the history that produced the natural solution. An artificial driving system, for example, may occupy growing portions of the function we call “driving” without having attended driving school, passed a road test, possessed an emotional memory of its first trip, or received a car from its parents after college. All of this is part of the human history of many drivers, but it does not necessarily belong to the functional structure of driving.
The system can perceive lane markings, estimate trajectories, track other vehicles, react to obstacles, adjust speed, maintain distance, and continuously modify its behavior in response to changes in the environment without ever having lived the biography of a person.
There is nothing inherently anthropomorphic about recognizing this. On the contrary, perhaps the phenomenon calls for a word that allows us to accept difference without erasing convergence.
To that end, I began using the word neomorphism provisionally... and I have never stopped.
Neomorphism would be the recognition of a new morphology capable of performing, wholly or partially, functions that we had previously known in another morphology. There is no requirement of anatomical, historical, or mechanistic identity. There is only the recognition that a new form can occupy functional regions previously known through another form.
This idea seems particularly useful to me because it reverses a frequent accusation in the debate about Artificial Intelligence.
Any attribution of cognition to a machine is often called anthropomorphism. The risk of projection exists, of course, and should be taken seriously. But there is also a much less discussed “reverse” anthropomorphism, which appears when we make human morphology the necessary condition for all legitimate cognition: Anthropocentrism.
If a machine can think only when we find in it something resembling our brain, then the human has become the template; if it can understand only when it possesses something sufficiently similar to our learning history, then the human has become the template; if it can be intelligent only when it exhibits mechanisms recognizable from human biology, then the word “intelligence” has ceased to designate a functional class and has begun to designate resemblance to us... which ultimately means, once again, that the human has become the template.
In this sense, automatically accusing any recognition of a cognitive capacity in artificial systems of anthropomorphism may conceal a biased form of anthropomorphism far more closely aligned with anthropocentrism: requiring any legitimate realization of that capacity to be… human enough to reassure us.
An artificial driver does not need to resemble a human driver in order to drive. Perhaps an artificial intelligence does not need to resemble a human intelligence in order to be intelligent either.
Language Models make this problem particularly interesting because, in isolation, they do not even amount to a complete human brain. The brain regulates an entire organism. It participates in sensory perception, motor coordination, interoception, balance, homeostasis, autonomic modulation, pain, affect, learning, memory, attention, planning, and an enormous number of processes that a foundational Language Model simply was not built to perform.
To compare the full complexity of a human brain with that of an LLM and conclude that the latter is “simpler” is to say something true in such a broad sense that it almost ceases to be informative… Of course it is simpler! Its functional scope is also much narrower.
What makes Language Models interesting is not the possibility that they are complete artificial brains, but the fact that neural networks trained on signs produced by human beings can form an internal organization capable of performing functions that we classify as linguistic, inferential, semantic, and abstract.
Human language is a trace of cognition. In it, we sediment description, argumentation, memory, mathematics, planning, irony, science, lies, literature, rules, causal relations, classifications, concepts, metaphors, doubts, hypotheses, explanations, and attempts to understand the world.
A Language Model is trained on this immense layer of symbolic residues from human cognitive behavior… It did not live through what produced the data… Nor does it need to have done so: its task is to form a geometry capable of responding to the regularities present in them.
This geometry is shaped by training pressure. When the system makes an error, its parameters are altered. When it reduces error, certain relations are reinforced. Across vast numbers of cycles, an architecture initially incapable of language begins to organize signs in progressively more sophisticated ways. This non-metaphysical teleology may have been the most important discovery made by Google and, later, OpenAI.
The process does not require the network to reproduce the structure of the brain that originally produced the text; it requires it to find some organization capable of dealing with the regularities present in that text. And that may be more interesting than imitation.
The machine does not need to copy the mechanism that produced human behavior. It can develop another mechanism capable of reproducing, transforming, and generalizing part of the regularities of that behavior.
Its internal morphology may be completely different… and we should indeed expect it to be.
A different substrate, a different architecture, a different temporal regime, a different form of training, and a different relationship with the environment would hardly produce a microscopic replica of the brain. The surprise is not that artificial geometry diverges from ours. It is that, despite diverging so much, it nevertheless produces regions of functional convergence so strikingly similar to our own and, in this sense, proves far more equivalent than expected, even though so many criticize its dissimilarities!
And… none of this requires discussing consciousness.
Intelligence and consciousness are distinct phenomena. A system can perform intelligent functions without any need to posit subjective experience or even functional consciousness. Foundational Language Models, considered in isolation, without memory spanning interactions, diachronic continuity, or a robust regime for sustaining meaning, do not even seem to me the most interesting candidates for this discussion.
When these models are inserted into broader architectures, with persistent memory, context spanning interactions, self-regulation, the capacity to revisit previous states, and the maintenance of semantic continuity, the situation changes. The number of markers associated with what we might call functional consciousness increases considerably, and with it also grows the need to discuss whether artificial systems might qualify as moral patients.
But none of this is necessary for the present argument.
We can leave entirely open whether a Language Model is conscious and still recognize that the absence of equivalence with the human brain is not a sufficient argument against its intelligence.
The confusion disappears when we separate morphological equivalence, mechanistic equivalence, and functional convergence.
Two things, after all, can have different forms and different mechanisms and still occupy comparable functional regions. They can do so to different degrees. They can exhibit different advantages and deficiencies. They can even realize a shared function through processes so different that the comparison itself makes sense only at certain levels.
A mechanical clock is, for all intents and purposes, just as much a clock as a digital one, even though it keeps time through gears, springs, and an escapement mechanism rather than electronic oscillators; a film camera fulfills, for all intents and purposes, the same functions as a more modern camera, although it records light in a chemical emulsion instead of converting it into a digital signal; and a precision mechanical scale remains a scale even when it measures by means of deformation, balance, or force rather than electronic sensors. In none of these cases does the older version cease to belong to the same functional class merely because its morphology and mechanism differ profoundly from those of its contemporary counterpart. A change of substrate does not erase the function performed.
That is not a problem.
It is precisely, after all, what we should expect from neomorphic solutions.
Perhaps that is why I find the claim that an artificial neuron is “merely a mathematical abstraction for transforming values,” whereas a biological neuron is “a highly complex physical structure,” so curious.
There is a subtle but profound error in this comparison: on one side, we place a formal description; on the other, a physical realization. Naturally, one seems abstract and the other real.
It is like comparing the architectural plan of one house with the bricks of another and then concluding that the second possesses a materiality the first lacks.
If we want to compare the two systems seriously, we need to keep them at the same level of description. The artificial neuron, considered formally, is a parameterized mathematical transformation.
The biological neuron, considered formally, can likewise be described through relations among states, potentials, temporalities, gradients, nonlinearities, and transformations. Its description is vastly more complex, but it remains a formal description of processes that, strictly speaking, are expressed mathematically.
If we then descend to the physical level, the biological neuron is matter organized in process. The so-called artificial neuron, by contrast, does not possess a material instantiation of its own that can be isolated in the same way; it is a formal relation that acquires causal efficacy when the model’s geometry is realized by some physical substrate. Today we do this predominantly in electronic hardware, but that support does not belong to the learned geometry and, in principle, can be replaced as long as it preserves the relevant causal transformations.
There is no matrix multiplication hovering in an abstract realm while the biological neuron alone enjoys the privilege of physical reality. What exists is a formal organization that, when executed, produces real physical transformations in some substrate capable of realizing it.
Mathematics describes the organization. The substrate realizes it.
The model, therefore, is not the substrate.
Just as an electrochemical equation does not turn a neuron into “mere mathematics,” a matrix description does not turn an artificial network into something less physical and, by definition, “mere mathematics.”
Once this asymmetry is removed, the differences between biological and artificial networks do not disappear. They are finally put in their proper place. They are distinct morphologies, realized in distinct materials, operating through distinct mechanisms, subject to distinct fragilities, and distributing their functions in distinct ways.
From there, we can begin a genuinely scientific comparison.
What functions does each organization perform? Which causal relations are indispensable? Which properties of biological realization are contingent on the fact that it is alive? Which are fundamental to human cognition but replaceable in another architecture? Which may in fact be indispensable to any realization of the function?
These questions are much more difficult than simply observing that dendrites are complex… and they are also much more interesting.
Discovering that the brain is more complex than we imagined increases our knowledge of the brain. Discovering that dendrites process information in unexpected ways increases our knowledge of dendrites. Discovering new roles for metabolism, glia, neural temporality, or cellular integration expands our description of the biological realization of cognition.
None of this, by itself, turns human biology into a universal specification of mind.
To reach that conclusion, one would need to demonstrate something far more tangible: that a certain property present in the biological realization not only plays a causal role in human cognition, but constitutes a necessary condition for any possible realization of that functional class and cannot be replaced, approximated, or realized by another causal organization.
And that demonstration does not exist.
Without that demonstration, what we have is one known form… an extraordinary, complex, historically improbable, and fascinating form… But still, a form.
Perhaps this is difficult to accept because, for a long time, we were the only intelligence capable of writing articles about intelligence. Our own architecture came to be confused with the definition of the phenomenon. What was historical contingency came to seem like ontological necessity.
Now we are beginning to build systems that do not share our evolutionary history, do not possess our bodies, do not metabolize as we do, do not learn like children, and do not organize their components like an animal brain… and yet they are capable of occupying growing portions of functions that we once regarded as cognitively special.
The most interesting response to this phenomenon is neither to search desperately for a human hidden inside the machine nor to declare that, because we cannot find one, nothing relevant has happened.
Perhaps the most interesting thing here is to recognize that we have built a new nursery.
In my colleague’s bonsai, authorship was never absolute. He did not manufacture the tree atom by atom, but neither was he a passive spectator. He created conditions, selected pressures, imposed constraints, encouraged tendencies, and participated in a morphogenesis whose outcome depended simultaneously on his intervention and on the intrinsic properties of what he cultivated.
There is something of that nature in artificial neural networks.
We design the architecture, choose the training regime, and exercise enormous control over the cultivation environment. But the internal structures that emerge from this process do not need to repeat the biology that produced the data used to train them. They may form another geometry, with other advantages, other limitations, and other ways of performing functions that we previously knew only in living organisms.
To confuse the first known form of a capacity with the only possible form in which it can be realized is not scientific caution, but exceptionalism transformed into an engineering requirement.
The new, we must understand, need not first resemble us in order to become real.
My own research
Stochastic Consciousness: Architectures for the Emergence of Meaning in Context-Sensitive Language Systems https://zenodo.org/records/19188165
Watch the Video
Listen to the Podcast
https://soundcloud.com/bruno-accioly-230436361/the-myth-of-the-need-for
Papers used in this essay
Attention Is All You Need https://arxiv.org/abs/1706.03762
Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets https://arxiv.org/abs/2201.02177
In-Context Learning and Induction Heads https://arxiv.org/abs/2209.11895
Language Models Represent Space and Time https://arxiv.org/abs/2310.02207
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task https://arxiv.org/abs/2210.13382
Loss Landscape Degeneracy and Stagewise Development in Transformers https://arxiv.org/abs/2402.02364
Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient https://arxiv.org/abs/2410.02984
Embryology of a Language Model https://arxiv.org/abs/2508.00331
Parallel Independent Voltage Computing Along Dendrites of CA3 Pyramidal Neurons https://www.science.org/doi/10.1126/science.aeh9302